A Multi-agent Reinforcement Learning Algorithm For Parking Resource Allocation Problem

As the economy continues thriving,traditional parking management methods cannot effectively use the follow-up parking information and can no longer meet the parking demands in the era of "Internet+".But still,the rapid development of the Internet and intelligent infrastructure make it possible to implement parking reservations.In this case,the city’s parking manager can allocate parking spaces forcibly for greater efficiency.However,the traditional optimization models of parking resource are all based on the reservation information which is acquired in advance.There are limited studies about the parking resource allocation under incomplete reservation information.In this context,this article proposes a dynamic allocation framework based on multiagent deep reinforcement learning with both reserved and non-reserved parking demands to achieve global optimization.It mainly consists of two parts: a simulation system that characterizes the operation of a parking lot,and a resource allocation algorithm based on QMIX.The latter will allocate a proper parking space to the reserved parking demand in real time by observing the operating information of the former,while the non-reserved demand selects an optimal parking space according to the User Equilibrium principle.By conducting numerical experiments on both typical and actual cases,the feasibility of the multi-agent resource allocation algorithm based on QMIX proposed in this paper is verified.And the typical experimental result also suggests that the QMIX algorithm shows a better allocation performance under a higher rate of reserved demand.Compared with existing management strategies,QMIX can bring a 4% improvement under 70% of the reservation rate and a 7% improvement under 90%.And in actual scenarios,it can save about 10% of total travel time.